nueva página del texto (beta)
Inglés (pdf)
Artículo en XML
Referencias del artículo
Traducción automática
Enviar artículo por email
Citado por SciELO 
Similares en
SciELO 
Permalink
1 Introduction
There has been an increase—in the last few years—in the number of LLMs available to the public, among such models we find: Pathways Language Model (PaLM) [4], BLOOM [28], Bidirectional Encoder Representations from Transformers (BERT) [6], BART [14], Robustly Optimized BERT Pretraining Approach (RoBERTa) [15], Generative Pre-trained Transformer (GPT) [23], GPT-2 [24], GPT-3 [3], and, more recently, Chat-GPT and GPT-4 [22].
Although great performance has been achieved in terms of text generation, there are some ethical issues that need to be addressed, for instance: the lack of validation for the data retrieved from these models, since they suffer from hallucinations—the information provided might be incorrect— [13]; the creation of fake news [34, 3] and academic cheating [1].
Since these models are publicly available, there exists high risk of misuse of LLMs models; therefore, there is a requirement for the creation of algorithms and systems capable of differentiating human generated text from that generated by LLMs. To this purpose, both, LLMs, as well as traditional Machine Learning (ML) have been tested on this classification task. When it comes to LLMs, several models have been used to leverage their capabilities (including BERT, RoBERTa, and GPT-2), while traditional ML have been less used (see Related work).
In this paper, we present our approach for the human vs. machine generated text classification problem, to this purpose, we selected the AuTexTification dataset [26] to work with. This dataset is divided, mainly, into two subtasks: 1) human vs. machine generated text and 2) model profiling, i.e., for a given machine generated sequence, determine which model created it.
Our proposal includes: a preprocessing stage, common in traditional NLP tasks, conformed by Tokenization, Multi-Word Token Expansion, Part of Speech and Lemmatization.
This stage is followed by Logistic Regression and Support Vector Machines (SVMs) models with 10-fold Cross Validation. Due to the particularities of both subtasks, special attention was taken into the feature extraction stages (previous to providing inputs to the models), creating combinations of unigrams, bigrams and trigrams of different nature, from PoS Taggins to Lemmas.
The rest of the paper is structured as follows: Section 2 provides a literature review on the present task, highlighting the most important research. Section 3 introduces the dataset, preprocessing and models employed for our study.
In Section 4 we present our empirical results from the experiments, as well as an overview of the meaning of such values. Lastly, Section 5 offers our conclusions drawn based on our results and observations.
2 Related Work
The task of detecting machine-generated text is closely related to and a direct consequence of the task of text generation, which was first implemented by authors such as Mann and McKeown [17, 18].
However, only recently has the problem of detecting whether a given text was generated or identifying the model that produced the text become challenging, even for humans [5]. Most efforts to detect or classify generated content apply Large Language Models such as BERT [10, 32], DistilBERT [21], RoBERTa [30, 7, 31, 12, 25, 9, 16, 19, 35], and others [2].
Classical approaches to detect generated text have also been attempted. Solaiman et al. [30] utilized classical machine learning classifiers such as logistic regression to identify content generated by GPT-2, finding that they do not perform worse than language models.
Unfortunately, this approach has not been completely followed by subsequent literature. In addition, statistical criteria have also been considered to address the task of detecting generated text [8, 20], as well as feature extraction [33, 29].
In the AuTexTification task [27], a preliminary evaluation of the task of detecting generated English text, logistic regression achieved a score of 0.6578 in F1-macro, surpassing the language model DeBERTa V3 (0.571), whereas in the model attribution task logistic regression reached a score of 0.3998, while DeBERTa V3 reached 0.6042.
3 Methodology
3.1 Dataset Description
The AuTexTification dataset [26] consists on data for two different subtasks: human vs machine generated text detection and model attribution. Therefore, this dataset contains text produced either by humans or by any of the following models: BLOOM-1b7, BLOOM-3b, BLOOM-7b1, GPT-3 Babbage, GPT-3 Curie and GPT-3 DaVinci-003.
Consequently, the training dataset for the first subtask consists on 33,845 text samples labeled as either machine-generated or human; on the other hand, for the second subtask, its training set includes 22,416 machine-generated text samples, each labeled with one of the letters A to F, corresponding to the modelo which generated the text.
The maximum text length is 98 and 97 words for subtasks one and two, respectively. Subtask 1 has a fixed testing set of 21,832 samples, whereas subtask 2 has 5,605 samples in its testing set. AuTexTification provides separate datasets for English and Spanish Languages, but our study only focuses on English. Figure 1 shows the structure of the dataset; it highlights in red the data used in this paper.
3.2 Preprocessing and Feature Extraction
3.2.1 Machine Learning Algorithms
We utilized a preprocessing pipeline for both tasks with the aim of extracting more significant syntactic and lexical features. This preprocessing sequence encompassed the following stages:
Tokenization: The raw text was broken down into individual tokens (words in this case).
Multi-Word Token (MWT) Expansion: This involved identifying combinations of words that function as a single unit and treating them as cohesive entities.
Part of Speech Tagging (PoS): To understand the grammatical roles and relationships within the text, we applied Part of Speech tagging. This involved assigning a specific tag, such as noun, verb, adjective, etc., to each token. Part of Speech tagging assists in capturing syntactic structures and provides valuable context for subsequent steps.
Lemmatization: Is a text normalization technique that involves reducing words to their base or root form.
The preprocessing steps applied to the raw text contribute significantly to its refinement, making it more suitable for analysis and improving classification accuracy. For our experiments, we included the sequence of lemmas and Part of Speech (PoS) tags as features.
The incorporation of PoS sequences was particularly important for both tasks, as it provided valuable insights into the writing style and enabled us to identify patterns and combinations of PoS tags that contributed to the text’s structure. Overall, utilizing PoS sequences enhanced our ability to extract syntactic information.
3.2.2 LSTM Ensemble
For this approach, we decided to work with two different PoS tags configurations, i.e., we used UPoS and XPoS, since the first provides a PoS analysis universal to most natural languages and the latter provides a specific PoS analysis for the given language which, in this particular case, is English.
The proposal was tested only on subtask 1 of the dataset (See Section 3.3), and involves the training of two LSTM cells, each one of them analysing—simultaneously— either UPoS or XPoS, where each tag was converted to a one-hot representation; once both LSTMs give their outputs, these are concatenated to become the input of a Multi-Layer Perceptron (MLP) of the following characteristics:
— A linear layer with a number of neurons equal to the sum of the hidden states of both LSTMs, for a total of 64 units.
— A dropout layer—for regularization purposes—, with a probability of 0.5 for every neuron to be zeroed out.
— A ReLU activation funciton layer, and,
— A single neuron as output, with a Sigmoid activation function for classification purposes.
Figure 2 shows the pipeline used for this ensemble proposal.
3.3 Subtask 1: Automatically Generated Text Identification
We deemed feature extraction crucial for this task, prompting us to experiment with various combinations of features:
a) Unigrams of lemmas with frequency counts.
b) Unigrams of lemmas with binary counts.
c) Unigrams of PoS with frecuency counts.
d) Unigrams of lemmas with binary counts along with unigrams of PoS and their frequency counts.
e) Unigrams, bigrams and trigrams of lemmas with binary counts.
f) Unigrams, bigrams and trigrams of PoS with frecuency counts.
g) Unigrams, bigrams and trigrams of lemmas with binary counts along with unigrams, bigrams and trigrams with their frecuency counts.
We employed Logistic Regression and Support Vector Machines (SVMs) for the classification, utilizing Stochastic Gradient Descent for both algorithms. In the case of the Support Vector Machine, a linear kernel was applied.
3.4 Task 2: Model Attribution
For this task, we experimented with the following features:
a) Unigrams of lemmas with binary counts.
b) Unigrams of PoS with frecuency counts.
c) Unigrams, bigrams and trigrams of lemmas with binary counts.
d) Unigrams, bigrams and trigrams of PoS with frecuency counts.
e) Unigrams, bigrams and trigrams of lemmas with binary counts along with unigrams, bigrams and trigrams with their frecuency counts.
Continuing with the previous proposal, we employed Logistic Regression and Support Vector Machines using Stochastic Gradient Descent.
4 Results
4.1 Validation Methods and Evaluation
As stated in Section 3.1, the AuTexTification tasks data is available for two subtasks, each one of them with four datasets: a training and a testing datasets, for English and Spanish Languages (see Figure 1).
For the purpose of this study, we exclusively focused on utilizing the training and testing data associated with subtasks one and two, for the English language.
We employed the training set as a development set. This involved partitioning the training data further to create a validation subset.
For the conventional machine learning techniques, we adopted a cross-validation approach as our chosen validation method.
The validation results presented in this study represent the average outcomes obtained from applying a 10-fold cross-validation process. On the other hand for our LSTM ensemble proposal, we used a hold out validation procedure on the training set, where 80% was still considered as training data, while the resting 20% became the validation set.
Figures 3 and 4 show a graphical representation of both 10-fold cross validation and hold out validation, respectively.
Subsequently, we proceeded to evaluate our models on the designated test set. The outcomes generated from this testing phase serve as the basis for our comparative analysis against both the state-of-the-art methods and the established baselines set forth in the contest.
4.2 Classification results. Subtask 1
4.2.1 LSTM ensemble
For the LSTM ensemble, the selected loss function was Binary Cross Entropy, with a learning rate of
To test the model’s performance, the corresponding testing set was used. This model only achieved an F1 average macro value of 0.62.
4.2.2 Machine Learning Algorithms
Initially, we tested the various features suggested in the preceding section using the development set. The results are presented in Table 1.
Table 1 Results of experiments of the subtask 1
| Model | Feat. | Recall | Precision | Accuracy | F1 score |
| LR | a | 0.7636 | 0.763 | 0.7633 | 0.0763 |
| b | 0.774 | 0.7735 | 0.7732 | 0.7731 | |
| c | 0.6215 | 0.5793 | 0.5811 | 0.53332 | |
| d | 0.7831 | 0.7807 | 0.7801 | 0.7797 | |
| e | 0.7405 | 0.7405 | 0.7405 | 0.7404 | |
| f | 0.761 | 0.7546 | 0.7538 | 0.7538 | |
| g | 0.8057 | 0.8055 | 0.8055 | 0.8054 | |
| SVM | a | 0.754 | 0.7552 | 0.7516 | 0.7512 |
| b | 0.7664 | 0.7616 | 0.7625 | 0.7587 | |
| c | 0.6425 | 0.6076 | 0.6063 | 0.573 | |
| d | 0.774 | 0.7616 | 0.7625 | 0.7587 | |
| e | 0.7301 | 0.7301 | 0.73 | 0.7299 | |
| f | 0.7558 | 0.7516 | 0.7518 | 0.7506 | |
| g | 0.8016 | 0.8014 | 0.8014 | 0.8013 |
The Table 1 highlights that the optimal feature combination for both classification algorithms is the combination ”g,” encompassing unigrams, bigrams, and trigrams of lemmas, as well as unigrams, bigrams, and trigrams for the PoS sequence.
As a result, we have chosen to employ these features for the test set.The results are presented in Table 2.
4.3 Classification results. Subtask 2
For subtask 2, we adhered to a similar approach, commencing with experimentation on the development set. The results are presented in Table 3.
Table 3 Results of experiments mentioned in the previous section
| Model | Feat. | Recall | Precision | Accuracy | F1 score |
| LR | a | 0.2897 | 0.2777 | 0.2793 | 0.2277 |
| b | 0.3844 | 0.3916 | 0.3924 | 0.3847 | |
| c | 0.4274 | 0.4399 | 0.4399 | 0.4302 | |
| d | 0.3332 | 0.3416 | 0.3415 | 0.3444 | |
| e | 0.459 | 0.4649 | 0.4646 | 0.4594 | |
| SVM | a | 0.2814 | 0.2513 | 0.2537 | 0.1945 |
| b | 0.3643 | 0.3701 | 0.3798 | 0.3647 | |
| c | 0.4182 | 0.4284 | 0.4284 | 0.4176 | |
| d | 0.3321 | 0.3208 | 0.3219 | 0.3088 | |
| e | 0.4585 | 0.4514 | 0.4517 | 0.4443 |
Table 3 demonstrates that the most effective feature blend for both classification algorithms is the ”e” combination.
This amalgamation comprises unigrams, bigrams, and trigrams of lemmas, along with unigrams, bigrams, and trigrams for the PoS sequence. Subsequently, we opted to utilize these features for the test set. The outcomes are outlined in Table 4.
5 Conclusions
We introduce a proposal addressing two distinct tasks that share similarities with the Natural Language Processing task of ”author attribution.”
Our approach places a stronger emphasis on feature extraction rather than on the model itself, employing traditional machine learning algorithms. Notably, both tasks surpass the baseline set by the AuTexTification contest organizers, utilizing Logistic Regression.
While our approach may not attain the pinnacle of state-of-the-art performance, it highlights the enduring importance of preprocessing and feature selection.
Tables 1 and 3 show us the significant impact of these factors, as they reveal substantial variations in outcomes even with the exact same model.
In the first task, our results with Support Vector Machine approach the state-of-the-art performance without resorting to large models, indicating that notable progress can be achieved. Furthermore, our findings underscore the pivotal role of syntactic information as a crucial feature when discerning authorship.
It is important to note that, although a Deep Learning (DL) architecture was proposed, we consider that the lower performance achieved by the model is inherent to the lack of a detailed text preprocessing, based strictly on the tasks to solve.
Therefore, special care must to be taken when proposing DL models, since the preprocessing must consider both, the characteristics of the text generation and model profiling tasks, as well as the type of preprocessing that works well with a given DL architecture.
We aspire to build upon our current approach by integrating these features into more sophisticated machine learning algorithms, such as Neural Networks. This progressive step aims to further enhance the quality of our outcomes.
